A Deep Learning-Based Intelligent Prediction Method and System for Spatiotemporal Distribution of PM2.5

Abstract

This invention proposes a deep learning-based intelligent prediction method and system for the spatiotemporal distribution of PM2.5. The method includes: collecting multi-source heterogeneous spatiotemporal datasets to construct a three-level nested hypergraph structure; using a multi-scale physical constraint module to construct a data-driven, physical law coupling layer; performing cross-scale feature fusion to obtain cross-scale fused features; quantifying uncertainty to obtain the prediction result distribution; then, based on this distribution and new multi-source heterogeneous spatiotemporal data, using an online incremental learning module for adaptive updating to obtain an adaptively updated model; and using an adversarial training enhancement module to generate adversarial examples, fusing them with the original training data, and inputting them into the model for iterative training to obtain an intelligent prediction model for the spatiotemporal distribution of PM2.5. This invention achieves deep coupling of multi-source heterogeneous spatiotemporal data through a three-level nested hypergraph structure, aggregates features at the functional area level supernodes, integrates data at the station level child nodes, and forms a cross-scale physical association network from pixel-level satellite data, thus solving the problem of local distortion in gridded systems.

Description

Technical Field

[0001] This invention relates to the field of PM2.5 prediction technology, and in particular to a deep learning-based intelligent prediction method and system for the spatiotemporal distribution of PM2.5. Background Technology

[0002] Currently, the field of PM2.5 spatiotemporal distribution prediction mainly relies on traditional statistical models, machine learning, and early deep learning frameworks. Traditional methods, such as the ARIMA time series model, focus on mining temporal patterns at single sites but struggle to capture spatial heterogeneity. Machine learning models (such as random forests and gradient boosting trees) integrate meteorological and land use data through feature engineering, but suffer from strong subjectivity in feature selection and weak cross-regional generalization ability. Early deep learning models, such as the CNN-LSTM combination, can process spatiotemporal data, but they often use gridded spatial representations, ignoring the physical boundaries and dynamic interaction features of urban functional areas (such as industrial and residential areas). Furthermore, the fusion of satellite remote sensing data (such as AOD) and ground monitoring data often remains at the pixel-level alignment level, lacking explicit modeling of multi-scale physical processes (such as atmospheric diffusion and turbulent mixing), leading to local distortions in prediction results in complex urban environments. Summary of the Invention

[0003] This invention aims to at least address the technical problem of prediction results being prone to local distortion in complex urban environments in existing technologies, and innovatively proposes a deep learning-based intelligent prediction method and system for the spatiotemporal distribution of PM2.5.

[0004] To achieve the above-mentioned objectives of this invention, this invention provides a deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5, the method comprising: S1. Collect multi-source heterogeneous spatiotemporal datasets to construct a three-level nested hypergraph structure, including functional area-level supernodes, site-level child nodes, and pixel-level satellite data; S2. Based on the three-level nested hypergraph structure, a data-driven, physical law coupling layer is constructed using a multi-scale physical constraint module; S3. Based on the data-driven, physical law coupling layer, cross-scale feature fusion is performed using a multi-scale feature pyramid to obtain cross-scale fused features. S4. Based on the cross-scale fusion features, the dynamic Bayesian inference module is used to quantify the uncertainty and obtain the distribution of the prediction results. S5. Based on the distribution of the prediction results and the new multi-source heterogeneous spatiotemporal data, an adaptive update model is obtained by using an online incremental learning module. S6. Based on the adaptive update model, the adversarial training enhancement module generates adversarial samples similar to the spatiotemporal distribution characteristics of real PM2.5. The adversarial samples are fused with the original training data and then input into the adaptive update model for iterative training. Finally, an intelligent prediction model for the spatiotemporal distribution of PM2.5 is obtained, and the prediction results are output.

[0005] In another aspect, the present invention also provides a deep learning-based intelligent prediction system for the spatiotemporal distribution of PM2.5, comprising: a processor and a memory for storing processor-executable instructions; and a deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5 implemented when the processor is configured to execute the executable instructions.

[0006] The beneficial effects of this invention are as follows: This invention achieves deep coupling of multi-source heterogeneous spatiotemporal data through a three-level nested hypergraph structure. Functional area-level supernodes dynamically aggregate spatial distance, functional area type, and meteorological similarity features; station-level subnodes integrate monitoring station and satellite pixel block data; and pixel-level satellite data, after bilinear interpolation alignment, forms a cross-scale physical association network, effectively solving the local distortion problem caused by neglecting the physical boundaries of urban functional areas in traditional gridded representations. Combined with a multi-scale physical constraint module, discretized atmospheric diffusion equations and learnable turbulent mixing coefficients are embedded into the hypergraph convolutional layer, achieving deep fusion of data-driven features and physical laws. At the microscale (100m), it captures details of industrial chimney emissions; at the mesoscale (1km), it tracks inter-city transmission paths; and at the macroscale (10km), it responds to the influence of monsoon / high-pressure systems. By dynamically balancing the weights of the three scales through a gated fusion unit, it significantly improves... The system improves prediction accuracy in complex urban environments. The dynamic Bayesian inference module fuses measured data with model predictions using Kalman filtering, decomposing cognitive and accidental uncertainty components to generate a prediction distribution containing confidence intervals, addressing the problem that traditional single-point estimation cannot support risk decision-making. The online incremental learning module uses cognitive uncertainty to guide gradient update step size and combines a memory replay mechanism to cache historical key samples, enabling adaptive updates of the model to the evolution of urban functional areas and changes in pollution sources, avoiding long-term prediction accuracy decay caused by offline retraining. The adversarial training enhancement module generates highly similar adversarial samples similar to the real distribution through a conditional generative adversarial network. The sample weights are adjusted according to the uncertainty components using a dynamic weighted fusion algorithm. After iterative training, the model's ability to fit complex PM2.5 spatiotemporal distribution characteristics is improved, ultimately achieving dual optimization of prediction accuracy and robustness on an independent validation set.

[0007] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0008] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of an intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning, according to the present invention. Detailed Implementation

[0009] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0010] Example 1 like Figure 1 As shown, a deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5 is proposed, the method comprising: S1. Collect multi-source heterogeneous spatiotemporal datasets to construct a three-level nested hypergraph structure, including functional area-level supernodes, site-level child nodes, and pixel-level satellite data; S2. Based on the three-level nested hypergraph structure, a data-driven, physical law coupling layer is constructed using a multi-scale physical constraint module; S3. Based on the data-driven, physical law coupling layer, cross-scale feature fusion is performed using a multi-scale feature pyramid to obtain cross-scale fused features. S4. Based on the cross-scale fusion features, the dynamic Bayesian inference module is used to quantify the uncertainty and obtain the distribution of the prediction results. S5. Based on the distribution of the prediction results and the new multi-source heterogeneous spatiotemporal data, an adaptive update model is obtained by using an online incremental learning module. S6. Based on the adaptive update model, the adversarial training enhancement module generates adversarial samples similar to the spatiotemporal distribution characteristics of real PM2.5. The adversarial samples are fused with the original training data and then input into the adaptive update model for iterative training. Finally, an intelligent prediction model for the spatiotemporal distribution of PM2.5 is obtained, and the prediction results are output.

[0011] The principle of the deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5 in this embodiment is as follows: By constructing a three-level nested hypergraph structure, multi-source heterogeneous spatiotemporal data is integrated to achieve deep coupling of urban functional areas, monitoring stations, and satellite remote sensing data. In the multi-scale physical constraint module, discretized atmospheric diffusion equations and learnable turbulent mixing coefficients are used to embed physical laws into the data-driven feature extraction process, thereby capturing the spatiotemporal evolution characteristics of PM2.5 concentration at different scales. Cross-scale feature fusion is completed through a multi-scale feature pyramid, dynamically balancing the weight distribution of micro, meso, and macro scales to improve the model's adaptability to complex environments. The dynamic Bayesian inference module, combined with Kalman filtering technology, effectively integrates measured data and model prediction results, decomposes cognitive uncertainty and accidental uncertainty, and provides confidence interval support for risk decision-making. The online incremental learning module guides gradient updates through cognitive uncertainty and combines a memory replay mechanism to cache key historical samples, ensuring that the model can adapt to the evolution of urban functional areas and changes in pollution sources, avoiding performance degradation caused by offline retraining. The adversarial training enhancement module generates highly similar adversarial samples and adjusts the sample weights through a dynamic weighted fusion algorithm to further optimize the model's ability to fit the complex spatiotemporal distribution characteristics of PM2.5, ultimately achieving a dual improvement in prediction accuracy and robustness.

[0012] As an optional embodiment of the present invention, optionally, constructing a three-level nested hypergraph structure includes: S101. Collect multi-source heterogeneous spatiotemporal datasets, and based on the multi-source heterogeneous spatiotemporal datasets, use geographic information data to divide the boundaries of urban functional areas, and generate functional area-level super nodes including industrial areas, residential areas and transportation hubs. Each super node dynamically aggregates three-dimensional features of spatial distance, functional area type and meteorological similarity through a hypergraph attention network. In step S101, it is necessary to explain in detail that the multi-source heterogeneous spatiotemporal dataset in this embodiment includes real-time monitoring data from ground stations (including pollutant concentrations and meteorological elements), satellite remote sensing data (such as MODIS / Aqua satellite AOD products), meteorological reanalysis data (such as ERA5 grid data), pollution source emission data (including industrial waste gas emissions), and geospatial basic data (such as urban road network vector data). These data cover different spatiotemporal resolutions: ground stations at the point scale, satellite remote sensing at the pixel scale, and meteorological reanalysis at the grid scale. After spatiotemporal alignment processing, they support the construction of a three-level nested hypergraph structure and feature aggregation.

[0013] Regarding the use of geographic information data to delineate urban functional zone boundaries: Based on high-resolution land use classification data, urban POI vector datasets, and VIIRS nighttime light radiation raster data, a multi-feature fusion spatial clustering method is adopted. First, the spatiotemporal alignment of each data source is preprocessed, POI data is aggregated into 100m×100m grid cells and POI density is calculated, and nighttime light data is resampled. Then, the core features of the grid cells are extracted, and the DBSCAN algorithm is used, incorporating similarity thresholds and spatial adjacency constraints for clustering. Finally, combined with manual verification and fine-tuning of the boundaries according to the urban master plan, a functional zone boundary vector layer is generated, with each functional zone corresponding to a supernode, and attributes including key information such as functional zone type.

[0014] S102. Based on the functional area-level supernode, embed site-level subnodes using geographic coordinate matching. The site-level subnodes include urban monitoring stations, meteorological stations, and satellite pixel blocks. In step S102, the following steps require detailed explanation: First, the latitude and longitude coordinates of urban PM2.5 monitoring stations (including national, provincial, and municipal levels) and ground meteorological stations are transformed using the WGS84 coordinate system. Then, the station coordinates are matched with the functional area boundary vector layer using a GIS spatial overlay analysis tool to determine the functional area-level supernode to which each station belongs. The hourly monitoring data (PM2.5 concentration, PM10 concentration, SO2 concentration, NOx concentration, temperature, humidity, wind speed and direction, etc.), station type (e.g., environmental quality monitoring point, traffic pollution monitoring point), and data quality labels (0-1 scores generated based on data missing rate and outlier percentage) are used as the core attributes of the station-level sub-nodes and associated with the corresponding functional area supernode. Second, for satellite pixel block data (e.g., MODISAOD10km×10km pixel blocks, Sentinel-5P...),... NO2 column concentration (7km×3.5km pixel block) was resampled to a uniform spatial resolution of 1km×1km using bilinear interpolation. Then, it was mapped to the functional area boundary through spatial grid indexing. Satellite pixel blocks covering the functional area were embedded as site-level child nodes into the corresponding supernodes. The child node attributes included the remote sensing inversion value of the pixel block, observation timestamp, satellite sensor type, and cloud cover. Finally, hyperedge connections were constructed between site-level child nodes and functional area supernodes. The hyperedge weights were dynamically calculated using a multilayer perceptron (MLP). The input features were the Euclidean distance between the site and the center of the functional area, the temporal continuity score of the site data, and the pollution source correlation between the functional area and the site (e.g., industrial monitoring sites in industrial areas have higher weights than residential sites). The weight values ​​were normalized to 0-1 to ensure that the child node features could participate in the feature aggregation process of the functional area supernodes according to their importance.

[0015] S103. Based on the site-level sub-nodes, the satellite remote sensing AOD pixel blocks are aligned with the site-level data using bilinear interpolation, and the pixel-level satellite data is integrated to obtain the pixel-level satellite data after bilinear interpolation alignment. In step S103, it is necessary to explain in detail that, firstly, for satellite aerosol optical depth (AOD) data (such as 10km×10km pixel blocks in MODISAqua / Terra), bilinear interpolation is performed with 1km×1km station-level sub-nodes within the functional area boundary as the target grid. A cloud cover constraint mechanism is introduced during interpolation. If the cloud cover of the original pixel block exceeds 30%, the interpolation results of adjacent valid pixel blocks are used to complete the data, ensuring the spatial continuity and validity of the data. The aligned 1km resolution AOD data is precisely matched with the station monitoring data (PM2.5 concentration, meteorological elements, etc.) within the corresponding grid using timestamps (time difference threshold of 15 minutes), and pixel-level and station-level sub-node hyperedge connections are constructed. The hyperedge weight is dynamically calculated by MLP based on the percentage of overlap area between the pixel block and the station, the data time synchronization score, and the satellite AOD inversion accuracy, and then normalized to 0-1.

[0016] Secondly, multi-source pixel-level satellite data is integrated. Sentinel-2 satellite 10m resolution surface reflectance data (blue, green, red, and near-infrared bands) and Sentinel-5P NO2 column concentration data (7km × 3.5km) are bilinearly interpolated to 1km resolution. These data, along with AOD interpolation results, constitute a pixel-level sub-node core attribute set, including spectral features, pollutant column concentration, and cloud cover. Pixel-level sub-nodes are then associated with corresponding functional area supernodes. The super-edge weights are dynamically adjusted based on functional area type, Euclidean distance between the pixel block and the functional area center, and data contribution, forming a cross-scale physical association network. This achieves deep coupling of multi-source heterogeneous spatiotemporal data from micro to macro levels.

[0017] Finally, quality control is performed on the aligned pixel-level satellite data. Outliers are filtered based on the correlation between the site-measured data and the interpolation results, and short-term fluctuations are handled using a time-series smoothing algorithm to ensure that the data meets the input requirements of the hypergraph convolutional layer. Through the above steps, pixel-level data embedding in a three-level nested hypergraph structure is completed.

[0018] S104. Based on the pixel-level satellite data after bilinear interpolation alignment, real-time meteorological conditions, human activity patterns and physical constraints, a multi-head attention mechanism is used to dynamically generate a hyperedge weight matrix to form a directed hyperedge connecting functional area-level supernodes and station-level subnodes. In step S104, it is necessary to explain in detail that the multi-head attention mechanism computes multiple attention heads in parallel to capture the interaction relationships of features in different subspaces. Each attention head independently calculates the association strength between functional area-level supernodes and station-level child nodes. The input includes real-time monitoring data of station-level child nodes (such as PM2.5 concentration and meteorological elements), pixel-level satellite data interpolation results (such as AOD and NO2 column concentrations), and aggregation attributes of functional area supernodes (such as functional area type and dominant pollution source). Physical constraints are introduced during the calculation, such as the pollutant transport direction and rate estimated based on the atmospheric diffusion equation, as prior guidance for attention weights to ensure that the hyperedge weight matrix conforms to physical laws. To enhance the model's adaptability to dynamic environmental changes, real-time meteorological conditions (such as wind speed, wind direction, and boundary layer height) and human activity patterns (such as traffic flow and industrial emission intensity) are encoded as time-series features, and long-term dependencies are extracted through gated recurrent units (GRUs). These time-series features are combined with static geographic information (such as functional area boundaries and station locations) to form a hybrid feature representation, which is input into the multi-head attention mechanism to optimize the dynamic generation process of hyperedge weights.

[0019] Finally, directed hyperedge connections are constructed using the generated hyperedge weight matrix to clarify the hierarchical relationship between functional area-level supernodes and site-level child nodes. Hyperedge weights reflect spatial distance, functional relevance, and dynamic changes over time. For example, at high wind speeds, the hyperedge weights of downwind sites increase significantly to reflect the spatial transport characteristics of pollutants. Simultaneously, the hyperedge weights are normalized to stabilize their values ​​between 0 and 1.

[0020] S105. Based on the functional area-level supernodes, site-level subnodes, and pixel-level satellite data, a three-level nested hypergraph structure is generated using the hypergraph structure storage module.

[0021] In step S105, it is necessary to explain in detail that, firstly, the hypergraph structure storage module achieves efficient organization and management of functional area-level supernodes, site-level child nodes, and pixel-level satellite data through a hierarchical design. Functional area-level supernodes store aggregated attributes and hyperedge connection information; site-level child nodes store real-time monitoring, interpolated satellite remote sensing data, and hyperedge weight matrices; pixel-level child nodes store core attribute sets and spatial mapping relationships.

[0022] To improve storage efficiency and query performance, a sparse matrix representation is used to compress and store the hyperedge weight matrix, reducing memory usage. A distributed storage architecture is introduced, with hypergraph data stored in shards to support parallel processing. A consistency check mechanism ensures data integrity and logical correctness. When updating the hypergraph structure, the hierarchical relationship and spatiotemporal alignment accuracy are checked; anomalies trigger a repair process. The hypergraph structure storage module in this embodiment employs a multi-layer indexing mechanism, combining geospatial and timestamp information to construct a global index table, enabling precise location retrieval. A built-in incremental update algorithm handles high-frequency data updates, avoiding full reconstruction. Graph embedding technology is introduced to reduce storage costs.

[0023] As an optional embodiment of the present invention, the construction of the data-driven, physical law coupling layer may include: S201. Based on the three-level nested hypergraph structure, the hypernode attribute matrix and hyperedge weight matrix are extracted using the multi-scale physical constraint module to generate an initial feature tensor containing spatial features and correlation information. The multi-scale physical constraint module analyzes the attributes and connections of functional area-level supernodes, site-level child nodes, and pixel-level child nodes in a three-level nested hypergraph structure, extracting the supernode attribute matrix and the hyperedge weight matrix. The supernode attribute matrix contains core features such as functional area type, while the hyperedge weight matrix reflects the spatial association, temporal variation, and interaction patterns between nodes. These matrices constitute the initial feature tensor, whose dimensions take into account spatial, temporal, and feature diversity, supporting subsequent analysis and modeling. When generating the initial feature tensor, multi-scale decomposition technology is introduced to process the data hierarchically. For data sources of different scales, wavelet transform or empirical mode decomposition is used to decompose the original signal into high-frequency and low-frequency components. High frequencies capture local details, while low frequencies represent global trends, thereby separating noise, highlighting key features, and improving information expression capabilities. To enhance the physical interpretability of the initial feature tensor, domain knowledge constraints are incorporated for optimization. For example, pollutant concentration distribution is corrected based on models such as atmospheric diffusion, and the hyperedge weight matrix is ​​smoothed using spatiotemporal consistency constraints. Through the above steps, the generated initial feature tensor has high-dimensional feature representation capabilities, reflecting the complex spatiotemporal and physical relationships of multi-source heterogeneous data.

[0024] S202. Based on the initial feature tensor, a differential equation constraint sublayer is constructed using the discretized atmospheric diffusion equation. The atmospheric diffusion equation is discretized using the finite difference method or the finite element method, transforming the continuous partial differential equation into a form suitable for discrete node calculations. Based on the functional area-level supernode distribution and station-level subnode monitoring data, regular or irregular grids are constructed to adapt to complex geographical environments. The grid is divided by combining the geometric characteristics of the functional area boundary vector layer to ensure clear and non-overlapping spatial correspondence between grid cells and supernodes. Then, the key variables of the atmospheric diffusion equation are mapped to the grid nodes, and their numerical distribution is initialized using real-time monitoring data and interpolation results from the initial feature tensor. When designing the differential equation constraint sublayer, a multi-scale time step strategy is introduced to balance computational accuracy and efficiency. Small time steps are used to solve the high-frequency fluctuating real-time monitoring data, while large time steps are used to reduce redundant calculations for low-frequency changing background field data. In addition, physical constraints, such as the law of conservation of mass, are embedded in the solution process to suppress the accumulation of numerical errors and improve the physical consistency of pollutant transport simulation. The output results of the differential equation constraint sublayer are integrated into the initial feature tensor to form an enhanced feature representation. The model weights the predicted pollutant concentrations and the initial feature tensor observations, dynamically adjusting the weights based on the uncertainty measures of both. If site-level sub-node monitoring data is missing or abnormal, the weight of the model's predicted values ​​is increased; otherwise, the original observation information is retained. This fusion mechanism utilizes prior knowledge from the physical model while preserving the flexibility of data-driven methods.

[0025] S203. Based on the differential equation constraint sublayer, a turbulent mixing model sublayer is constructed using a learnable turbulent mixing coefficient. The mixing coefficient is automatically adjusted through backpropagation to adapt to the pollutant mixing effect at different scales. Learnable turbulent mixing coefficients are introduced using deep learning parameterization methods, replacing the fixed coefficients of traditional turbulence models with trainable variables. Backpropagation is used for optimization during training to adapt to complex turbulent characteristics. Specifically, the initial state of the turbulent mixing model sub-layer is constructed based on the pollutant concentration field and wind speed field distribution constrained by differential equations. The spatiotemporal distribution of the turbulent mixing coefficients is dynamically adjusted in conjunction with real-time meteorological conditions.

[0026] In the sub-layer design of the turbulence mixing model, a multi-scale feature fusion technique is employed to capture turbulence effects at different resolutions. For example, local convolution operations are used to extract features from high-resolution satellite data, while global pooling operations are used to integrate the average turbulence intensity for low-resolution functional area-level supernode attributes. This approach not only addresses the challenge of data resolution differences but also enhances the model's sensitivity to local turbulence anomalies.

[0027] To ensure physical consistency, an observation-based correction mechanism is embedded, using real-time monitoring data from station-level sub-nodes to perform posterior correction on the turbulent mixing coefficient, thus avoiding physical inconsistencies. Simultaneously, a sparse regularization term is introduced to constrain the learning process, preventing overfitting and improving generalization ability. Through these steps, the turbulent mixing model sublayer can accurately simulate the mixing and diffusion behavior of pollutants at different scales.

[0028] S204. Based on the differential equation constraint sublayer and the turbulent mixing model sublayer, a data-driven feature and physical law deep fusion is performed using a hypergraph convolutional layer cascade structure to generate a data-driven and physical law coupling layer containing a physical constraint regularization term weight matrix, a learnable turbulent mixing coefficient matrix, and a fused feature tensor.

[0029] The cascaded hypergraph convolutional layer structure consists of multiple stacked multi-head hypergraph convolutional layers, each modeling cross-scale feature interactions of pixel-level, site-level, and functional area-level hypernodes. The enhanced feature tensor output from the pre-sequence differential equation constraint sublayer and the learnable turbulent mixing coefficient matrix generated by the turbulent mixing model sublayer are used as initial inputs to the first hypergraph convolutional layer. This layer performs aggregation operations on the hypernode attribute matrix and the hyperedge weight matrix, while simultaneously transforming physical constraints into regularization loss terms. The weights of these regularization loss terms are dynamically allocated by the physical constraint regularization weight matrix, assigning higher weights to functional area-level hypernodes and appropriately reducing the weights to pixel-level child nodes.

[0030] The multi-layer hypergraph convolutional layer employs a bidirectional cross-scale feature transfer mechanism, where lower-layer output features are passed upwards and upper-layer global physical constraint information is passed downwards, achieving deep coupling between microscopic and macroscopic aspects. Each layer introduces a multi-head hypergraph attention mechanism to capture the association patterns between nodes in different subspaces in parallel.

[0031] The fusion layer at the end of the cascaded structure concatenates the hypergraph convolutional output features from each layer, and generates the final fused feature tensor through adaptive dimensionality reduction. This tensor integrates data-driven spatiotemporal features, physical constraints, and learnable turbulent mixing effects, along with a weight matrix for physical constraint regularization and a learnable turbulent mixing coefficient matrix. During backpropagation, the physical constraint regularization guides the model parameter updates, ensuring that the fused features conform to atmospheric physical laws. Through these steps, a data-driven, physically coupled layer with both data fitting capabilities and physical consistency is formed.

[0032] As an optional embodiment of the present invention, obtaining cross-scale fusion features may include: S301. Based on the data-driven and physical law coupling layer, local convolution kernels are used to extract local pollution source features at the microscale, generating a micro feature map containing spatial details. Local convolutional kernel design needs to be customized and optimized based on the spatial morphology and emission characteristics of micro-pollutant sources. For point sources such as industrial chimneys, a 3×3×3 three-dimensional local convolutional kernel is used, whose large receptive field can capture the initial morphology of pollutant diffusion. For line sources such as traffic intersections, a combination of 1×3 and 3×1 convolutional kernels is designed to extract pollutant concentration gradient features. For area sources such as construction sites, a large 5×5 convolutional kernel is selected to cover the pollution diffusion range. During feature extraction, the hotspot areas of the pollution source (cloud cover <10% and AOD value >0.8) are first located from the fused feature tensor, and then the initial micro-feature map is output by sliding calculation using a customized convolutional kernel.

[0033] To enhance the information representation capability of the initial microscopic feature map, a multi-dimensional feature enhancement mechanism is introduced. On the one hand, the initial feature map is nonlinearly activated, and the ReLU6 function is used to suppress abnormally high values. Batch normalization is used to stabilize the feature distribution. On the other hand, a spatial attention module is embedded, and an attention weight matrix is ​​calculated based on real-time meteorological elements. High weights are assigned to feature areas upwind of the pollution source to highlight the dominant transmission path, while downwind and non-pollution source areas are weighted to reduce redundancy.

[0034] Consistency verification ensures the accuracy and reliability of the micro-feature map. It is then compared with the spatial annotation information in the ground pollution source census database, and the Intersection over Union (IoU) is calculated. If the IoU value is lower than a preset threshold (e.g., 0.75), convolution kernel parameter adjustments or attention module weight updates are triggered; if the requirements are met, the micro-feature map is used as the basic input for subsequent cross-scale fusion, and its feature dimensions include core attributes such as the pollution source type probability vector.

[0035] S302. Based on the microscopic feature map, a global convolutional kernel is used to capture the transmission trend features of the mesoscopic scale region to generate a mesoscopic trend feature map. The global convolutional kernel design needs to adapt to the transport characteristics of mesoscale regions (1-10km spatial range). A 9×9×5 spatiotemporal global convolutional kernel is adopted, whose large receptive field can capture the transport path and trend of pollutants. The kernel parameters are initialized by pre-training with the regional meteorological background field, and different initial weights are assigned for different transport scenarios, and the distribution of convolutional weights is optimized. During processing, the microscopic feature map generated by S301 is first interpolated to mesoscale resolution, and then the global convolutional kernel is used for operation. At the same time, an adaptive meteorological weight modulation module is introduced to dynamically adjust the weight coefficients according to real-time meteorological data.

[0036] To enhance the temporal dynamic representation capability of the mesoscopic trend feature map, a spatiotemporal sliding window aggregation mechanism is introduced. This mechanism constructs a spatiotemporal window to calculate statistical features and integrates them into the initial feature map. Simultaneously, a regional transport constraint module is embedded to verify pollutant flux based on the law of mass conservation. If the deviation exceeds a threshold, the spatial weight distribution of the convolutional kernel is adjusted.

[0037] The reliability of the feature map is ensured by spatial autocorrelation verification. The Moran's I index of each site-level child node is calculated to evaluate the degree of spatial clustering. If the index is less than 0.5, the convolution kernel parameters are fine-tuned. If the requirements are met, a meso-level trend feature map containing the core attributes is output.

[0038] S303. Based on the meso-scale trend feature map, the macro-scale spatial aggregation range is adjusted using an adaptive convolution kernel to extract the influence features of large-scale meteorological models on pollutant distribution and generate a macro-meteorological correlation feature map. The design of adaptive convolutional kernels requires flexible adjustments based on the spatial scale and dynamic characteristics of large-scale meteorological systems. The core is to construct a dual-drive mechanism of "circulation perception - scale adaptation." For different types of large-scale circulation (such as extratropical cyclones, subtropical highs, and monsoon troughs), three basic convolutional kernel sizes are preset: 15×15×24 (covering a single city and surrounding areas, 24-hour time dimension), 25×25×48 (covering urban clusters, 48-hour time dimension), and 35×35×72 (covering inter-provincial areas, 72-hour time dimension). The convolutional kernel parameters are initialized using pre-training based on large-scale meteorological field characteristics in ERA5 reanalysis data (such as 500hPa geopotential height, 700hPa relative humidity, and near-surface wind speed). For example, for stable scenarios under subtropical high control, the weight of the convolutional kernel's time dimension is strengthened to capture pollutant accumulation trends; for strong wind scenarios of extratropical cyclones, the weight of the spatial dimension is optimized to highlight long-distance transport paths. During the processing, the current large-scale circulation type is first automatically determined by the circulation type identification model (a meteorological field classifier based on CNN-LSTM). Then, the basic convolution kernel is dynamically selected and its edge weights are fine-tuned according to the type. For example, when it is identified as the outer circulation of a typhoon, the weight of the outer region of the basic 25×25 convolution kernel is increased by 1.3 times to capture the inhibitory effect of the typhoon's outer sinking airflow on pollutants.

[0039] To enhance the cross-scale correlation capability of macro-meteorological correlation feature maps, a meteorological field-pollutant residual fusion mechanism is introduced. This mechanism residually connects large-scale circulation features (such as the westerly wave phase) in the macro-feature map with regional transport flux in the meso-trend feature map. After dimensionality matching is achieved through a 1×1 convolutional layer, a weighted sum of the two is calculated (the weights are dynamically determined by the correlation coefficient between circulation intensity and regional transport flux). This preserves the dominant role of the macro-meteorological field without losing meso-scale transport details. Simultaneously, a long-distance transport contribution calculation module is embedded. Based on the meteorological field gradient and pollutant concentration gradient in the macro-feature map, it quantifies the contribution ratio of different large-scale circulations to PM2.5 concentration in the target area (e.g., 35% for westerly transport and 65% for local accumulation), and integrates this ratio as an additional channel into the feature map.

[0040] S304. Based on the three-scale feature maps—microscopic feature map, mesoscopic trend feature map, and macroscopic meteorological correlation feature map—a gated fusion unit is used to dynamically calculate the weight coefficients of each scale. A spatial attention mechanism is used to balance the contribution of local details and global trends, generating a cross-scale fused feature tensor containing multi-scale spatial features and physical constraint information. The expression for dynamically calculating the weight coefficients of each scale using the gated fusion unit is as follows: in, Indicates spatial location Normalized weighting coefficients at micro / meso / macro scales Indicates spatial location The channel dimension Softmax function, This represents the convolution kernel of the gated fusion unit. This represents the convolution operation. Representing the spatial location of microscale feature maps The feature vector (100m resolution, including local details such as industrial chimney emissions). Representing the spatial location of mesoscale feature maps The feature vector (1km resolution, including regional trends such as inter-city pollutant transport paths). Representing the spatial location of macroscopic scale feature maps The feature vector (10km resolution, including the influence of large-scale meteorological models such as monsoon / high pressure systems). Indicates the spatial location of the fused feature map eigenvectors, Indicates spatial location Eigenvectors at micro / meso / macro scales It represents the Hadamardi (or Hadama) stack.

[0041] The initialization of convolutional kernel parameters in the gated fusion unit incorporates prior knowledge of multi-scale feature contribution. For example, during periods of high PM2.5 pollution, the macro-meteorological correlation feature map is assigned an initial weight of 0.4, the meso-trend feature map 0.35, and the micro-feature map 0.25. Subsequent weight adjustments are made dynamically through backpropagation to adapt to different scenarios. A spatial attention mechanism integrates the spatial distribution characteristics of real-time meteorological elements. For spatial location i, the channel dimension Softmax function incorporates meteorologically sensitive factors such as boundary layer height as additional input channels. When the boundary layer height is below 500m, the weight coefficient of local pollution source areas in the micro-feature map is increased by 1.1 times; when it is above 1500m, the weight coefficient of the macro-meteorological correlation feature map is increased by 1.2 times. To address the inconsistent resolution of multi-scale feature maps (100m for micro-scale, 1km for meso-scale, and 10km for macro-scale), before calculating the Hadamard product, an adaptive sampling module interpolates the meso- and macro-scale feature maps to the micro-scale resolution. Linear interpolation is used for prevailing wind areas, and cubic spline interpolation is used for stable meteorological conditions. Meteorological weight modulation is embedded in the interpolation to reduce errors. To ensure the reliability of the fused features, a multi-scale consistency verification mechanism is introduced: First, a local-global feature correlation verification is performed, calculating the Pearson correlation coefficient of concentration trends in corresponding micro and macro regions. If the coefficient is below 0.6, the weights of the gated fusion units are fine-tuned. Second, a physical mass conservation verification is performed, adjusting the spatial attention mechanism weights within ±10% of the relative error of the total mass of each functional region between the input emission inventory and the fused feature tensor. The final output is a cross-scale fused feature tensor with 128 channels, integrating core attributes from the micro (32 channels), meso (48 channels), and macro (48 channels) perspectives. This tensor can be directly input into the prediction head module for predicting the spatiotemporal distribution of PM2.5 concentration.

[0042] As an optional embodiment of the present invention, obtaining the distribution of prediction results may include: S401. Based on the cross-scale fusion feature tensor, the initial state distribution of dynamic Bayesian inference is initialized using the state space model to generate an initial state vector containing spatiotemporal dependencies. The state-space model employs an Extended Kalman Filter (EKF) architecture to adapt to the nonlinear dependence of PM2.5 spatiotemporal distribution. The initial state vector's dimension is consistent with the number of channels in the cross-scale fusion feature tensor (128 dimensions), with each dimension corresponding to a core attribute in the fusion feature (such as micro-level pollution source emission intensity, meso-level functional zone transport flux, and macro-level meteorological field sensitivity coefficient). During initialization, combining historical PM2.5 concentration monitoring data from the same season and time period over the past 14 days with the statistical distribution of cross-scale features, values ​​are assigned to the state variables at each spatial location: the mean concentration is taken as 1.1 times the historical mean (to adapt to the real-time nature of the current fusion feature), and the variance is taken as 1.2 times the historical variance (to preserve the uncertainty space), with physical constraints (such as concentration value ≥ 0) applied. The initial weight design of the state transition matrix incorporates spatiotemporal dependency priors: In the spatial dimension, the transition probability of adjacent spatial locations follows an exponential distance decay law (distance coefficient k=0.75), and the transition probability of non-adjacent regions is based on the weight coefficient of the mesoscopic transmission path; in the temporal dimension, the transition probability adopts the AR(2) autoregressive model coefficients (the coefficient is 0.8 for 1-hour lag and 0.15 for 2-hour lag) to capture the concentration continuity in a short period of time. At the same time, a rationality verification mechanism for the initial state distribution is introduced: the KL divergence between the initial state distribution and the historical state distribution is calculated. If the KL divergence exceeds a preset threshold (such as 0.3), the parameter fine-tuning process of the state transition matrix is ​​triggered (based on gradient descent to optimize the transition weights) to ensure that the initial state vector can accurately reflect the spatiotemporal dependency relationship.

[0043] S402. Based on the initial state vector, a nonlinear state transition sublayer is constructed using the discretized atmospheric diffusion equation. The spatiotemporal evolution consistency of pollutant concentration gradient and wind field direction is constrained by Gaussian process prior, thereby generating a state transition matrix that conforms to physical law constraints. The expression for the discretized atmospheric diffusion equation is: in, Indicates spatial location In time PM2.5 concentration values, This indicates that the time is long since the walk. Indicates spatial location In time PM2.5 concentration values, Indicates the spatial position in the x-direction in time. The wind speed component, and Indicates the distance between the space and the walking distance. Indicates spatial position in the y direction In time The wind speed component, This represents the learnable turbulent mixing coefficient. Indicates spatial location In time The pollution sources are strong. The x-direction concentration gradient is represented by the central difference method (discretized to reflect the physical constraint that "the pollutant concentration gradient is consistent with the wind direction"). The above expression uses the explicit finite difference method to discretize the atmospheric diffusion equation. The left side is the concentration prediction value at the next moment, and the right side includes the convection term, diffusion term, and source term. The initial value of the learnable turbulent mixing coefficient K is calculated based on the turbulence intensity data (such as friction velocity u) observed at the local meteorological station (the formula is K=0.4u²z / L, where z is the observation height and L is the Moning-Obukhov length). Subsequent backpropagation dynamic optimization is used to adapt to diffusion characteristics under different meteorological scenarios such as static stability and strong winds. The pollution source strength S(x,y,t) is directly extracted from local source terms such as industrial emissions and traffic emissions in the microscopic feature map, ensuring consistency between the source terms and the actual emission inventory. The core of the Gaussian process prior constraint is to use the consistency between the concentration gradient and the wind field direction as a constraint term of the covariance function: for any spatial location (i,j) and time t, if the angle θ between the concentration gradient direction and the wind field direction exceeds 30°, the probability density of this state is reduced by adjusting the weights of the covariance matrix, forcing adherence to the physical law of "pollutants being transported along the dominant wind field direction." The covariance function expression is C(θ)=exp(-θ² / (2 152)) The smaller θ is, the higher the reliability. The nonlinear state transition sublayer embeds the discretized equation into the state update process: first, the sum of the convection, diffusion, and source terms is calculated, and then the state vector is updated by combining the Gaussian process constraint results; a residual correction mechanism is introduced, if the residual between the updated state vector and the mesoscopic transport flux exceeds ±20%, then the weights of K or S(x,y,t) are adjusted. When constructing the state transition matrix, sparsity constraints are applied, retaining only the non-zero transition probabilities of adjacent spatial locations (distance ≤2km) and continuous time steps (≤3 hours); after generation, it needs to be verified: first, the sum of the transition probabilities is close to 1 (error ≤±5%), otherwise it is normalized; second, the evolution trend is consistent with the direction of the macroscopic wind field, otherwise the weights of the corresponding elements are fine-tuned, and finally a state transition matrix that conforms to the physical laws is obtained.

[0044] S403. Based on the state transition matrix and the measured PM2.5 data from the monitoring stations, a Gaussian observation model is constructed using the Bayesian update rule. The measured data and model predictions are fused using Kalman filtering, and the posterior distribution of the state is dynamically adjusted to generate an intermediate prediction distribution that includes cognitive uncertainty and accidental uncertainty. In this embodiment, the method for constructing the Bayesian-updated Gaussian observation model is as follows: PM2.5 data measured at monitoring stations are used as observations, and the observed values ​​are fused with the model's predicted values ​​using a Kalman filter framework. Specifically, when initializing the observation matrix of the Gaussian observation model, the principle of adapting to station distribution density and spatial resolution is considered: for areas covered by microscopic feature maps (100m resolution), if monitoring stations exist, their observations are directly mapped to their corresponding spatial locations; for areas covered by mesoscopic and macroscopic feature maps (1km and 10km resolutions), a weighted average method is used to integrate data from multiple stations, with weights determined by an inverse distance function to ensure the spatial representativeness of the observations. The design of the observation noise covariance matrix R needs to consider station instrument errors and environmental interference factors. For example, in high humidity environments (relative humidity > 80%), the observation noise variance is increased (by a factor of 1.5) for stations susceptible to wet deposition to reduce the impact of measurement bias caused by wet deposition on the posterior distribution.

[0045] To address the separation of cognitive uncertainty and random uncertainty in modeling, this embodiment proposes a dual-channel uncertainty quantification strategy: cognitive uncertainty mainly stems from the incompleteness of model parameters and structure, and is estimated by analyzing the distribution characteristics of historical prediction residuals; random uncertainty originates from observation noise and random perturbations, and is quantified based on the observation noise covariance matrix R. The two types of uncertainty are modeled separately using Gaussian processes and then fused in a weighted manner in the final prediction distribution, with the weights dynamically determined by the signal-to-noise ratio in the training data.

[0046] S404. Based on the intermediate prediction distribution, the total uncertainty is decomposed by Monte Carlo Dropout sampling and quantile regression, generating cognitive uncertainty components and accidental uncertainty components, and capturing time-dependent features through a long short-term memory network. The core idea of ​​Monte Carlo Dropout sampling is to estimate the total uncertainty by randomly discarding some neurons during the prediction phase to simulate the posterior distribution of model parameters. Specifically, during each forward propagation, a portion of the cross-scale fused feature tensor is randomly masked with probability p (e.g., 0.2), and sampling is repeated N times (e.g., N=50) to obtain N independent distributions of prediction results. Quantile regression is used to decompose cognitive uncertainty from random uncertainty in the sampling results: by calculating different quantiles (e.g., 0.05, 0.5, 0.95 quantiles) for the N sets of prediction values, the range of random uncertainty is quantified; simultaneously, cognitive uncertainty is assessed based on the distribution of the deviation between the predicted mean and the true value. The introduction of Long Short-Term Memory (LSTM) networks aims to capture the time-dependent characteristics of PM2.5 concentration changes, with its input being the uncertainty component sequence processed by quantile regression. The hidden layer dimension of the LSTM is set to 64, and the time step covers historical data from the past 12 hours, ensuring that it can learn short-term concentration fluctuation patterns and long-term trend evolution. Furthermore, an attention mechanism is added to the LSTM output to dynamically adjust the importance weights of inputs at different times to the current prediction. For example, higher weights are assigned to periods of sudden pollution events (increased by 1.3 times) to enhance the model's responsiveness to abnormal events. The resulting cognitive uncertainty and accidental uncertainty components will serve as important inputs to subsequent prediction head modules to further optimize the prediction accuracy and reliability of the spatiotemporal distribution of PM2.5 concentration.

[0047] S405. Based on the cognitive uncertainty component and the accidental uncertainty component, a Gaussian mixture model is used to fit the distribution of the prediction results, generating a distribution tensor of the prediction results containing the mean, variance and confidence interval.

[0048] The construction of the Gaussian mixture model needs to consider the multi-peak characteristics of PM2.5 concentration distribution and the statistical regularity of uncertainty components. Specifically, the number of mixture components is set to 3 during model initialization, corresponding to low-pollution, medium-pollution, and high-pollution scenarios, respectively. The initial weights of each component are determined based on the frequency of occurrence of different pollution levels in historical data. The estimation of the mean vector integrates the weighted results of cognitive uncertainty components and accidental uncertainty components. The design of the covariance matrix introduces spatiotemporal dependency constraints: in the spatial dimension, the covariance values ​​of adjacent locations follow an exponential decay law (decay coefficient λ=0.8), while non-adjacent areas are adjusted based on the weight coefficients of the mesoscopic transport path; in the temporal dimension, the covariance values ​​capture the concentration continuity in a short period of time through autoregressive model coefficients (0.7 for 1-hour lag and 0.2 for 2-hour lag). The confidence interval is generated using the quantile regression method, calculating the 95% confidence interval by using the cumulative distribution function of the predicted results distribution to ensure that the interval coverage can reflect the fluctuation characteristics of the actual observed data. In addition, to improve the model's adaptability to extreme pollution events, an additional mixing component is added for high-pollution scenarios, and the parameter estimation is optimized through the expectation-maximization algorithm. The final output prediction result distribution tensor contains mean, variance, and confidence interval information.

[0049] As an optional embodiment of the present invention, obtaining the adaptive update model may include: S501. Based on the distribution of the prediction results and the newly acquired multi-source heterogeneous spatiotemporal data, the new satellite remote sensing AOD pixel blocks and station-level real-time data are aligned using bilinear interpolation to generate a new spatiotemporally aligned dataset. The newly acquired multi-source heterogeneous spatiotemporal data includes satellite remote sensing aerosol optical depth (AOD) data, real-time PM2.5 concentration data from ground monitoring stations, and dynamically updated meteorological field data. To ensure data alignment accuracy, bilinear interpolation was optimized using a local terrain correction factor: for complex terrain areas (such as mountainous or coastal areas), elevation difference weights were introduced to adjust the interpolation coefficients, avoiding error amplification caused by terrain undulations; for temporal misalignment, a sliding window mechanism was used to integrate the dynamic trends over the past 3 hours to smooth the impact of short-term fluctuations. The generated spatiotemporally aligned dataset must meet the consistency verification standard, i.e., the correlation between the AOD value at any spatial location and the corresponding PM2.5 concentration should be higher than 0.7; otherwise, a resampling process is triggered. Furthermore, the spatial resolution of the dataset is uniformly set to a 1km × 1km grid, and the temporal resolution is 1 hour, facilitating subsequent model processing.

[0050] S502. Based on the new dataset, dynamically update the site-level child node attributes in the three-level nested hypergraph structure using the hypergraph structure storage module to generate the updated three-level nested hypergraph structure. The three-level nested hypergraph structure comprises three levels: a micro-level (site-level child nodes), a meso-level (functional area nodes), and a macro-level (meteorological field nodes). Site-level child node attributes include real-time PM2.5 concentration, observation noise variance, spatial representativeness weight, and environmental interference markers. During dynamic updates, the real-time observation values, instrument calibration status, and current environmental parameters (such as relative humidity and wind speed) of the corresponding stations are first extracted from the new dataset. The real-time concentration value directly replaces the original attributes of the child nodes. If the instrument is uncalibrated, the observation noise variance is increased to 1.3 times its original value. The spatial representativeness weight is recalculated based on the inverse proportionality function of the distance between the station and the surrounding grid (distance coefficient k=0.6). The environmental interference marker is set to "high interference" or "normal" based on the humidity threshold (>80%). Subsequently, the edge weights of the station child nodes and the mesoscopic functional area parent nodes are updated synchronously: based on the real-time emission contribution rate of the functional area where the station is located (output by the microscopic pollution source feature extraction module), the edge weight increases by 0.12 for every 10% increase in the contribution rate; the edge weights of the station child nodes and the macroscopic meteorological field parent nodes are dynamically optimized according to the wind field's influence coefficient on the station (the cosine of the angle between the wind field direction and the pollutant transport path of the station), with a higher weight for a smaller angle (maximum increase of 0.25). After the update is completed, a hypergraph structure consistency check is performed: first, the coefficient of variation of the attributes of all station child nodes in the microscopic layer is calculated. If it exceeds a preset threshold (e.g., 0.4), the attributes of abnormal stations are smoothed (based on the mean of neighboring stations); second, the aggregated concentration of the mesoscopic functional area nodes (obtained by the weighted average of the subordinate station child nodes) is verified to be consistent with the predicted trend of the macroscopic meteorological field nodes. If the deviation exceeds 15%, the edge weights of the corresponding level are fine-tuned; finally, a stable and attribute-accurate updated three-level nested hypergraph structure is generated.

[0051] S503. Based on the updated three-level nested hypergraph structure, cognitive uncertainty component and accidental uncertainty component, the gradient update step size of the online incremental learning module is dynamically adjusted using an adaptive learning rate algorithm to generate an uncertainty-guided gradient update strategy. The adaptive learning rate algorithm uses the AdamW optimizer as its basic framework and combines cognitive uncertainty (Uc) and accidental uncertainty (Ua) to construct a two-layer weight guidance mechanism. First, the total uncertainty guidance weight w = α is defined. Uc+β Ua, where α is initialized to 0.6 (prioritizing correction of cognitive uncertainty) and β is initialized to 0.4 (balancing noise suppression of accidental uncertainty); the base learning rate η0 is set to 1e-4, and the dynamic learning rate η = η0. (1+0.5 Uc-0.3 For higher cognitive uncertainty (larger Uc), the learning rate is increased to accelerate parameter optimization; for higher random uncertainty (larger Ua), the learning rate is reduced to avoid noise interference.

[0052] The adjustment of the gradient update step size needs to be considered in relation to the hierarchical characteristics of the three-level nested hypergraph: the gradient step size coefficient for micro-level (site-level) nodes is 1.2 (to adapt to high-frequency real-time data updates), the coefficient for meso-level (functional area) nodes is 0.9 (to maintain the stability of regional aggregation characteristics), and the coefficient for macro-level (meteorological field) nodes is 0.8 (to mitigate over-adjustment of the slowly changing characteristics of the meteorological field). The sliding window mean of the uncertainty components is calculated every 5 time steps (i.e., 5 hours), and the values ​​of w and η are updated synchronously. Simultaneously, a gradient norm constraint is introduced: if the current gradient norm exceeds 1.0, it is multiplied by a decay factor of 0.8 to prevent gradient explosion and ensure the numerical stability of the update process. The resulting gradient update strategy ensures both rapid adaptation to high-uncertainty regions and maintains the overall stability and generalization ability of the model.

[0053] S504. Based on the gradient update strategy guided by the uncertainty, historical key samples are cached through the memory replay mechanism, an incremental training set is constructed by combining the new dataset, and the new and old features are fused by the hypergraph convolutional layer to generate an incremental feature tensor that fuses physical constraints. The core of the memory replay mechanism is to cache historical samples that are crucial for model updates. These include samples from sudden pollution events (such as PM2.5 concentrations rising by more than 50 μg / m³ within one hour), samples with historical prediction errors exceeding 30%, and typical spatiotemporal samples covering low, medium, and high pollution levels. The cache capacity is the key samples from the last 7 days. When new key samples are added, the oldest non-key samples are replaced using a FIFO strategy. Simultaneously, cached samples are weighted, with sudden pollution event samples having a weight of 1.5 and high-error samples having a weight of 1.2, ensuring that high-value samples are used first. The incremental training set is constructed using a hybrid mode of "new data + replay samples" in a 7:3 ratio. Replay samples are generated through weighted random sampling to balance the model's adaptability to new dynamics with the continuity of historical patterns.

[0054] When fusing new and old features in the hypergraph convolutional layer, the updated three-level nested hypergraph structure is first transformed into a hyperadjacency matrix. Feature propagation is then performed using the hypergraph node attribute vectors. Micro-level node features are aggregated with meso-level functional region nodes, and the meso-level features then interact with macro-level meteorological field nodes. A physical constraint regularization term is introduced during propagation to avoid violating atmospheric diffusion laws. Next, the feature tensors of the new dataset and historical replay samples are input into different branches. The branch weights are dynamically adjusted using an attention mechanism. The initial weight of the new data branch is 0.6, and the weight of the historical sample branch is 0.4. If the cognitive uncertainty of a certain region in the new data exceeds the mean by 20%, the weight of the new data branch for that region is increased to 0.75. The hypergraph convolutional layer consists of two layers, each with a hidden dimension of 128. The activation function is ReLU, and L2 regularization is added to prevent overfitting.

[0055] The final generated incremental feature tensor with fused physical constraints has a spatial resolution of 1km×1km and a temporal resolution of 1 hour. The feature dimensions include multiple cross-scale fused features, providing comprehensive spatiotemporal-physical fusion input for the subsequent prediction head module and supporting the model's high-precision prediction.

[0056] S505. Based on the incremental feature tensor of the fused physical constraints, the weight matrix of the physical constraint regularization term is dynamically adjusted using Bayesian update rules to generate an adaptive update model.

[0057] The core of the Bayesian update rule is to dynamically update the posterior distribution of the weight matrix of the physical constraint regularization terms based on the deviation information between the actual observations and model predictions in the incremental feature tensor, thereby achieving adaptive synergy between physical constraints and the data-driven model. The specific operation is as follows: First, define the prior distribution of the weight matrix of the physical constraint regularization terms. The initial weight matrix prior is set as a Dirichlet distribution, and the parameter α0 corresponds to the initial importance allocation of each physical constraint term, ensuring coverage of key physical mechanisms of atmospheric diffusion. Next, calculate the likelihood function to quantify the explanatory power of different physical constraint terms for the deviation based on the prediction deviation in the incremental feature tensor. Then, update the posterior distribution of the weight matrix using the Bayesian formula, and take the mean as the new weight matrix.

[0058] During the dynamic adjustment process, weight allocation is optimized for different scenarios. For example, for samples of sudden pollution events, the posterior weight of the pollution source emission contribution constraint is increased, and the posterior weight of the terrain elevation influence constraint is adjusted in mountainous areas. After adjustment, the effect needs to be verified by calculating the model's prediction confidence interval coverage and mean absolute error reduction rate on the incremental training set. If the coverage is insufficient, the weight of meteorological field-related constraints is increased; if the mean absolute error reduction does not meet the target, the weights of pollution source or terrain constraints are fine-tuned.

[0059] Finally, the updated physical constraint regularization term weight matrix is ​​embedded into the loss function of the prediction head module, guiding the model to learn spatiotemporal distribution characteristics that conform to atmospheric physical laws. The adaptively updated model can maintain historical prediction accuracy, quickly adapt to new spatiotemporal dynamics, and output the spatiotemporal distribution prediction results of PM2.5 concentration.

[0060] As an optional embodiment of the present invention, optionally, obtaining the intelligent prediction model for the spatiotemporal distribution of PM2.5 includes: S601. Based on the adaptive update model, a conditional generative adversarial network is used to generate adversarial samples similar to the real distribution of PM2.5 using the real spatiotemporal distribution characteristics as conditions. A discriminator network is used in combination with a gradient penalty mechanism to align the feature distribution of the adversarial samples with that of the real samples, thereby generating a set of highly similar adversarial samples. The generator of the Conditional Generative Adversarial Network (CGAN) employs a spatio-temporal U-Net structure. The input layer contains an adaptively updated predicted mean tensor, cognitive / accidental uncertainty components, and physical constraint feature vectors. The intermediate layers embed a Spatio-Temporal Attention (STA) module to capture the spatio-temporal dependence of PM2.5 concentration: the spatial attention branch extracts spatial correlation weights between grid cells through 3×3 convolutional layers, while the temporal attention branch utilizes a Transformer encoder (containing four attention heads) to learn hourly dynamic variations. The generator output is a tensor of adversarial examples with a 1km×1km grid and 1-hour resolution, with dimensions consistent with real samples (containing PM2.5 concentration values, uncertainty markers, and physical feature labels).

[0061] The discriminator employs a spatiotemporal PatchGAN structure. The input is a concatenated tensor of adversarial examples or real samples and conditional features (conditional features include real-time weather dynamics, pollution source emission intensity, and terrain elevation attributes). The output is the true / false probability of a local spatiotemporal region (e.g., a 3×3 grid with a continuous 3-hour window). To avoid pattern collapse and stabilize the training process, the discriminator introduces a gradient penalty mechanism based on Wasserstein distance: the gradient norm of the discriminator with respect to the input samples is calculated; if the norm deviates from 1, a penalty term is added to the loss function to ensure that the discriminator satisfies the Lipschitz continuity constraint.

[0062] The generation of adversarial examples requires verification based on physical constraints: in addition to the adversarial loss, the generator's loss function incorporates a consistent physical constraint regularization term (dynamically updated weight matrix). For example, the cosine value of the angle between the wind direction and the pollutant transport path in the adversarial example must be ≥0.6; otherwise, fine-tuning of the generator parameters is triggered. The generated adversarial examples undergo double verification: first, a spatiotemporal feature cosine similarity ≥0.85 with the real examples; second, a KL divergence ≤0.2. If these conditions are not met, the generator's learning rate (base η=5e-5, increasing by 10% each time the condition is not met) or the discriminator's penalty coefficient is adjusted. The final set of high-similarity adversarial examples comprises 30% of the original training data and covers low, medium, and high pollution levels, as well as different terrain types such as plains, mountains, and coastlines, ensuring the generalization of adversarial training.

[0063] S602. Based on the adversarial example set and the original training data, a dynamic weighted fusion algorithm is used to mix them according to a preset ratio. The spatiotemporal coordinates and functional area attributes of the fused data are aligned through the hypergraph structure storage module to generate a spatiotemporally consistent fusion training dataset. The expression of the dynamic weighted fusion algorithm is: in, Indicates spatiotemporal location The fused feature vector Indicates spatiotemporal location The cognitive uncertainty component, Indicates spatiotemporal location The random uncertainty component, Indicates spatiotemporal location Adversarial sample feature vectors Indicates spatiotemporal location The original training data feature vector.

[0064] The core of the dynamic weighted fusion algorithm lies in adjusting the contribution ratio of adversarial examples to the original training data through uncertainty components, thereby balancing the model's focus on high-uncertainty regions and its stability in low-noise regions during the fusion process. Specifically, the cognitive uncertainty component (Uc) and the random uncertainty component (Ua) are transformed into weight coefficients through nonlinear mapping functions: for cognitive uncertainty, a Sigmoid function is used to map it to the interval [0.5, 1] ​​to ensure that adversarial example features in high-uncertainty regions receive higher weights; for random uncertainty, an inverse proportional function is used to map it to the interval [0, 0.5] to suppress the influence of adversarial examples in areas with strong noise interference. In addition, to avoid data imbalance caused by extreme weights, upper and lower limits for the weights are set to 0.9 and 0.1, respectively.

[0065] The fused feature vectors need to be further aligned with spatiotemporal coordinates and functional area attributes through a hypergraph structure storage module. The hypergraph structure design includes three layers: the micro-layer corresponds to 1km×1km grid cells, the meso-layer corresponds to functional area aggregation nodes, and the macro-layer corresponds to global meteorological field features. The alignment process first maps the fused feature vectors to hypergraph nodes based on spatiotemporal location indices. Then, nodes are classified and labeled using functional area attribute labels (such as industrial areas, residential areas, and nature reserves), and nodes with similar attributes are connected by hyperedges to enhance the consistency of features between regions. At the same time, a time sliding window mechanism is introduced to smooth the fused feature vectors over three consecutive hours, reducing feature abrupt changes caused by time discontinuities.

[0066] S603. Based on the fused training dataset, the gradient descent algorithm is used to dynamically adjust and adaptively update the model parameters to generate an iterative model; The gradient descent algorithm employs the Adam optimizer with an initial learning rate of 1e-4, combined with a dynamic learning rate adjustment strategy to improve training efficiency and model convergence. Specifically, the dynamic learning rate adjustment is based on the rate of change of loss on the validation set: if the validation loss decreases by less than 0.001 over three consecutive epochs, the learning rate is reduced to 0.8 times its original value; if the validation loss shows an upward trend, an early stopping mechanism is immediately triggered to avoid overfitting. Simultaneously, gradient clipping is introduced to prevent gradient explosion, with a gradient norm threshold set to 5 to ensure the stability of model parameter updates.

[0067] During the iteration process, the spatiotemporal consistency features of the fused training dataset are further propagated and optimized through hypergraph convolutional layers. Micro-level node features first aggregate information from adjacent grid cells through hyperedges, forming local spatial dependencies; meso-level functional region nodes utilize global pooling to integrate features from all micro-nodes within the region, extracting representative regional patterns; macro-level meteorological field nodes capture dynamic patterns in long-term sequences through temporal convolutional networks, enhancing the model's adaptability to changes in meteorological conditions. L1 regularization terms (coefficient 1e-6) are added to features at each level during propagation to sparsify less important feature connections, improving the model's interpretability and generalization performance.

[0068] S604. Based on the distribution of prediction results of the iterative model on the independent validation set, the hyperparameters are dynamically adjusted using the Bayesian optimization algorithm to generate the final intelligent prediction model for the spatiotemporal distribution of PM2.5. In step S604, it is necessary to explain in detail that the core of the Bayesian optimization algorithm lies in constructing a Gaussian process surrogate model to efficiently explore and utilize the hyperparameter space, thereby minimizing the prediction error on the validation set. Specifically, firstly, the hyperparameter search space is defined: the learning rate range is set to [1e-5, 1e-3], the candidate batch size is [32, 64, 128], the candidate hidden dimensions of the hypergraph convolutional layer are [64, 128, 256], and the candidate number of heads for the attention mechanism is [2, 4, 8]. Next, the Gaussian process surrogate model is initialized, and its objective function value (e.g., mean absolute error MAE) on the validation set is calculated based on the initial sampling points (e.g., randomly selecting 5 sets of hyperparameter combinations).

[0069] In each iteration, the next hyperparameter combination to be evaluated is selected through a collection function (e.g., the desired improvement in EI), and applied to the training and validation process of the iterative model. If the objective function value of the new combination is better than the current optimal value, the surrogate model is updated and the optimal hyperparameter configuration is recorded. Simultaneously, a dynamic confidence interval adjustment strategy is introduced: when the validation error fluctuates significantly, the variance of the noise term in the Gaussian process is increased to enhance robustness; when the validation error tends to stabilize, the variance of the noise term is decreased to accelerate convergence.

[0070] S605. Based on the final PM2.5 spatiotemporal distribution intelligent prediction model, output the predicted PM2.5 concentration at future spatiotemporal scales through model inference.

[0071] The input for model inference needs to integrate multi-source real-time and historical data, including real-time meteorological observation data (wind speed, wind direction, temperature, relative humidity, etc., with a sampling frequency of once per hour) at a 1km×1km grid scale, online pollution source monitoring data (industrial emissions, vehicle traffic, etc., with an update interval of 5 minutes), digital elevation model (DEM) data, and the spatiotemporal sequence of PM2.5 concentration over the past 7 days. The input data is first aligned in spatiotemporal coordinates and matched in attributes through a hypergraph structure storage module to ensure consistency between micro-grid units, meso-functional zones, and macro-meteorological field characteristics.

[0072] During inference, the input features are propagated hierarchically through the hypergraph convolutional layers: the micro-level grid features aggregate spatial correlation information of adjacent units through hyperedges, the meso-level functional area features integrate the differentiated contributions of industrial / residential / natural areas within the region, and the macro-level meteorological field features capture future meteorological trends through temporal convolution. Subsequently, the prediction head module, which integrates the weight matrix of physical constraint regularization terms, generates the core prediction results and calculates the satisfaction of each physical constraint term (such as the cosine similarity between wind speed direction and diffusion path, the Pearson correlation coefficient between pollution source emissions and concentration distribution, etc.).

[0073] The prediction results contain multi-dimensional information: ① Core prediction value: the average PM2.5 concentration of the next 24 / 48 / 72 hours for a 1km×1km grid; ② Uncertainty quantification: the prediction variance of each grid cell (reflecting the model's cognitive uncertainty about the area) and the 95% confidence interval, etc.

[0074] Example 2 A deep learning-based intelligent prediction system for the spatiotemporal distribution of PM2.5 includes: a processor and a memory for storing processor-executable instructions; and a deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5 implemented when the processor is configured to execute the executable instructions.

[0075] It should be noted that the computer device includes a processor and memory, and may also include one or more of the following: multimedia components, input / output (I / O) interfaces, and communication components. The processor controls the overall operation of the computer device, completing all or part of the steps of the deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5. The memory stores various types of data to support device operation and can be implemented by volatile or non-volatile storage devices or combinations thereof, such as static random access memory.

[0076] The multimedia component includes a screen (such as a touch screen) and an audio component, the audio component being used to output and / or input audio signals, including a microphone to receive external audio signals, and also including at least one speaker to output audio signals.

[0077] I / O interfaces provide interfaces for the processor and other interface modules, such as keyboards, mice, virtual or physical buttons.

[0078] Communication components enable wired or wireless communication between computer devices and other devices. Wireless communication includes Wi-Fi, Bluetooth, etc., and corresponding communication components include Wi-Fi modules, etc.

[0079] As a preferred option, the computer device can be implemented using electronic components such as application-specific integrated circuits to execute a deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5.

[0080] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5, characterized in that, The method includes: A three-level nested hypergraph structure was constructed by collecting multi-source heterogeneous spatiotemporal datasets, including functional area-level supernodes, site-level child nodes, and pixel-level satellite data; Based on the aforementioned three-level nested hypergraph structure, a data-driven, physical law-coupled layer is constructed using a multi-scale physical constraint module. Based on the data-driven, physical law coupling layer, cross-scale feature fusion is performed using a multi-scale feature pyramid to obtain cross-scale fused features. Based on the cross-scale fusion features, the dynamic Bayesian inference module is used to quantify uncertainty and obtain the distribution of prediction results. Based on the distribution of the prediction results and the new multi-source heterogeneous spatiotemporal data, an adaptive update model is obtained by using an online incremental learning module for adaptive updating. Based on the adaptive update model, the adversarial training enhancement module generates adversarial samples similar to the spatiotemporal distribution characteristics of real PM2.

5. The adversarial samples are then fused with the original training data and input into the adaptive update model for iterative training. Finally, an intelligent prediction model for the spatiotemporal distribution of PM2.5 is obtained, and the prediction results are output.

2. The intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning as described in claim 1, characterized in that, Constructing a three-level nested hypergraph structure includes: Collect multi-source heterogeneous spatiotemporal datasets, and based on these datasets, use geographic information data to delineate the boundaries of urban functional areas, generating functional area-level supernodes that include industrial areas, residential areas, and transportation hubs. Each supernode dynamically aggregates three-dimensional features—spatial distance, functional area type, and meteorological similarity—through a hypergraph attention network. Based on the functional area-level supernode, site-level subnodes are embedded using geographic coordinate matching. The site-level subnodes include urban monitoring stations, meteorological stations, and satellite pixel blocks. Based on the site-level sub-nodes, the satellite remote sensing AOD pixel blocks are aligned with the site-level data using bilinear interpolation, and the pixel-level satellite data is integrated to obtain the pixel-level satellite data after bilinear interpolation alignment. Based on the pixel-level satellite data aligned by bilinear interpolation, real-time meteorological conditions, human activity patterns and physical constraints, a multi-head attention mechanism is used to dynamically generate a hyperedge weight matrix, forming a directed hyperedge connecting functional area-level supernodes and station-level subnodes. Based on the functional area-level supernodes, site-level subnodes, and pixel-level satellite data, a three-level nested hypergraph structure is generated using the hypergraph structure storage module.

3. The intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning as described in claim 1, characterized in that, Building a data-driven, physical law-coupled layer includes: Based on the three-level nested hypergraph structure, the hypernode attribute matrix and hyperedge weight matrix are extracted using the multi-scale physical constraint module to generate an initial feature tensor containing spatial features and correlation information. Based on the initial feature tensor, a differential equation constrained sublayer is constructed using the discretized atmospheric diffusion equation; Based on the differential equation constraint sublayer, a turbulent mixing model sublayer is constructed using learnable turbulent mixing coefficients. The mixing coefficients are automatically adjusted through backpropagation to adapt to the pollutant mixing effects at different scales. Based on the differential equation constraint sublayer and the turbulent mixing model sublayer, a data-driven feature and physical law deep fusion is performed using a hypergraph convolutional layer cascade structure, generating a data-driven and physical law coupling layer containing a physical constraint regularization term weight matrix, a learnable turbulent mixing coefficient matrix, and a fused feature tensor.

4. The intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning as described in claim 1, characterized in that, Obtaining cross-scale fusion features includes: Based on the data-driven, physical law-coupled layer, local convolutional kernels are used to extract local pollution source features at the microscale, generating a micro-feature map containing spatial details. Based on the micro-feature map, a global convolutional kernel is used to capture the transmission trend features in the meso-scale region, generating a meso-scale trend feature map. Based on the meso-scale trend feature map, the macro-scale spatial aggregation range is adjusted by using an adaptive convolution kernel to extract the influence features of large-scale meteorological models on pollutant distribution and generate a macro-meteorological correlation feature map. Based on the three-scale feature maps—microscopic feature map, mesoscopic trend feature map, and macroscopic meteorological correlation feature map—the weight coefficients of each scale are dynamically calculated using a gated fusion unit. The contribution of local details to the global trend is balanced through a spatial attention mechanism, generating a cross-scale fusion feature tensor that contains multi-scale spatial features and physical constraint information.

5. The intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning as described in claim 4, characterized in that, The expression for calculating the weight coefficients for each scale is as follows: in, Indicates spatial location Normalized weighting coefficients at micro / meso / macro scales Indicates spatial location The channel dimension Softmax function, This represents the convolution kernel of the gated fusion unit. This represents the convolution operation. Representing the spatial location of microscale feature maps eigenvectors, Representing the spatial location of mesoscale feature maps eigenvectors, Representing the spatial location of macroscopic scale feature maps eigenvectors, Indicates the spatial location of the fused feature map eigenvectors, Indicates spatial location Eigenvectors at micro / meso / macro scales It represents the Hadamardi (or Hadama) stack.

6. The intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning as described in claim 1, characterized in that, The distribution of the predicted results includes: Based on the cross-scale fusion feature tensor, the initial state distribution of dynamic Bayesian inference is initialized using a state-space model to generate an initial state vector containing spatiotemporal dependencies. Based on the initial state vector, a nonlinear state transition sublayer is constructed using the discretized atmospheric diffusion equation. The spatiotemporal evolution consistency of pollutant concentration gradient and wind field direction is constrained by Gaussian process prior, thereby generating a state transition matrix that conforms to physical constraints. Based on the state transition matrix and the measured PM2.5 data from the monitoring stations, a Gaussian observation model is constructed using Bayesian update rules. The measured data and model predictions are fused using Kalman filtering, and the posterior distribution of the state is dynamically adjusted to generate an intermediate prediction distribution that includes cognitive uncertainty and accidental uncertainty. Based on the intermediate prediction distribution, Monte Carlo Dropout sampling and quantile regression are used to decompose the total uncertainty, generating cognitive uncertainty components and accidental uncertainty components, and capturing time-dependent features through a long short-term memory network. Based on the cognitive uncertainty component and the random uncertainty component, a Gaussian mixture model is used to fit the distribution of the prediction results, generating a distribution tensor of the prediction results containing the mean, variance and confidence interval.

7. The intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning as described in claim 1, characterized in that, Obtaining an adaptive update model includes: Based on the distribution of the prediction results and the newly acquired multi-source heterogeneous spatiotemporal data, the new satellite remote sensing AOD pixel blocks and station-level real-time data are aligned using bilinear interpolation to generate a new spatiotemporally aligned dataset. Based on the new dataset, the attributes of the site-level child nodes in the three-level nested hypergraph structure are dynamically updated using the hypergraph structure storage module to generate the updated three-level nested hypergraph structure. Based on the updated three-level nested hypergraph structure, cognitive uncertainty component and accidental uncertainty component, the gradient update step size of the online incremental learning module is dynamically adjusted using an adaptive learning rate algorithm to generate an uncertainty-guided gradient update strategy. Based on the uncertainty-guided gradient update strategy, historical key samples are cached through a memory replay mechanism, an incremental training set is constructed by combining the new dataset, and new and old features are fused using hypergraph convolutional layers to generate an incremental feature tensor that incorporates physical constraints. Based on the incremental feature tensor of the fused physical constraints, the weight matrix of the physical constraint regularization term is dynamically adjusted using Bayesian update rules to generate an adaptive update model.

8. The intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning as described in claim 1, characterized in that, The intelligent prediction model for the spatiotemporal distribution of PM2.5 includes: Based on the aforementioned adaptive update model, a conditional generative adversarial network is used with the spatiotemporal distribution characteristics of real PM2.5 as a condition. The generator network generates adversarial samples similar to the real distribution, and the discriminator network combined with the gradient penalty mechanism is used to align the feature distribution of the adversarial samples with that of the real samples, thereby generating a set of highly similar adversarial samples. Based on the adversarial sample set and the original training data, a dynamic weighted fusion algorithm is used to mix them at a preset ratio. The spatiotemporal coordinates and functional area attributes of the fused data are aligned through the hypergraph structure storage module to generate a spatiotemporally consistent fusion training dataset. Based on the fused training dataset, the gradient descent algorithm is used to dynamically adjust and adaptively update the model parameters to generate an iterative model. Based on the distribution of prediction results of the iterative model on the independent validation set, the hyperparameters are dynamically adjusted using the Bayesian optimization algorithm to generate the final intelligent prediction model for the spatiotemporal distribution of PM2.

5. Based on the final PM2.5 spatiotemporal distribution intelligent prediction model, the model inference outputs the predicted PM2.5 concentration at future spatiotemporal scales.

9. The intelligent prediction method for the spatiotemporal distribution of PM2.5 based on deep learning as described in claim 8, characterized in that, The expression for the dynamic weighted fusion algorithm is: in, Indicates spatiotemporal location The fused feature vector Indicates spatiotemporal location The cognitive uncertainty component, Indicates spatiotemporal location The random uncertainty component, Indicates spatiotemporal location Adversarial sample feature vectors Indicates spatiotemporal location The original training data feature vector.

10. A deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5, characterized in that, The system includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement the deep learning-based intelligent prediction method for the spatiotemporal distribution of PM2.5 as described in any one of claims 1 to 9 when executing the executable instructions.

Images